Dynamic Facial Expression Generation on Hilbert Hypersphere With Conditional Wasserstein Generative Adversarial Nets.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generating dynamic facial expressions from static images is challenging.
• Previous methods often struggle with realistic motion, appearance, and identity preservation.

Purpose of the Study:

• To propose a novel approach for generating videos of six basic facial expressions from a neutral face image.
• To leverage manifold-valued representations and generative adversarial networks (GANs) for dynamic facial expression synthesis.

Main Methods:

• Modeling facial landmark motion as curves on a hypersphere.
• Utilizing a conditional manifold-valued Wasserstein GAN for motion generation.
• Employing a second conditional GAN for texture editing to generate image sequences.

Main Results:

• Generated realistic videos with continuous motion, appearance, and preserved identity.
• Demonstrated effectiveness on Oulu-CASIA and MUG Facial Expression datasets.
• Achieved state-of-the-art results in dynamic facial expression generation.

Conclusions:

• The proposed manifold-valued GAN approach is effective for dynamic facial expression generation.
• The framework supports expression transfer and data augmentation for emotion recognition.
• This work pioneers the use of manifold-valued representations with GANs for this task.